Popular AI assistants process users’ conversations, but their policies differ on what they retain, how they use it, and what controls they provide.
We reviewed the privacy policies, data practices, and opt-out mechanisms for eight popular AI assistants: ChatGPT, Gemini, Claude, Copilot, Meta AI, DeepSeek, Perplexity, and Grok. The results paint a picture of an industry where privacy protections vary wildly, defaults almost always favor the company, and the apps built by the biggest tech companies tend to be the worst offenders.
The Rankings: From Least to Most Invasive
Incogni’s 2026 AI Privacy Ranking evaluated platforms across three categories: how they use data for model training, how transparent they are about it, and how much data their apps collect and share. The results largely track with a separate analysis by Captain Compliance.
Here’s how the major platforms stack up:
Least invasive:
- Le Chat (Mistral AI) - Collected and shared the least data of any platform tested. Limited data collection, strong opt-out options, and the mobile apps are minimal in what they request.
- ChatGPT (OpenAI) - The most transparent about its practices. Clear privacy policy, straightforward opt-out for model training, and explicit documentation about what gets stored. That said, ChatGPT now shows ads on free and Go tiers, which means conversation topics inform ad targeting.
- Grok (xAI) - Decent opt-out functionality but loses points on transparency and the breadth of data it collects from X (formerly Twitter) interactions.
Middle of the pack:
- Claude (Anthropic) - Anthropic announced a change in August 2025 for Free, Pro, and Max users. Those users had until September 28 to accept updated terms, and leaving the “You can help improve Claude” setting enabled allowed data from new and resumed chats to be used for model improvement. Anthropic said data shared under that setting could be retained for five years. Work, Gov, Education, and API customers were not affected.
- Copilot (Microsoft) - Enterprise versions come with no-train commitments, SOC 2 Type 2 certification, and HIPAA compliance. The consumer version is a different story. Concentric AI found that Copilot accessed nearly three million sensitive records per organization in the first half of 2025 - and in February 2026, Microsoft confirmed a bug that let Copilot bypass DLP policies and read confidential emails.
Most invasive:
- Perplexity - CEO Aravind Srinivas has been explicit about wanting more user data, including building a browser specifically to track users outside the app. Shares mobile identifiers, hashed emails, and cookie data with advertisers. Uses prompts for training on all plans. Messages and user files were stored without encryption.
- DeepSeek - Stores all data on servers in China. Collects keystroke patterns, device data, and IP addresses. The iOS app sent device information without encryption. Security researchers found code with built-in capability to send data to the Chinese government. Wiz Research documented a publicly accessible database containing over a million lines of chat histories. Multiple countries have banned or restricted it.
- Meta AI - Ranked dead last in Incogni’s evaluation. The review found that Meta AI collected precise locations and addresses, shared data with third parties, and offered no opt-out mechanism for model training. SCWorld also noted that Meta uses a single privacy policy across its products.
What They All Have in Common
A Stanford HAI study examined all six major AI companies (Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI) and found a consistent pattern: all of them use chat data by default to train their models, and some keep that data indefinitely.
The researchers also flagged a practice specific to multi-product companies like Google, Meta, Microsoft, and Amazon: user interactions get merged with data from their other products. Your search queries, purchase history, social media engagement, and AI conversations all get stitched together into a single profile.
Privacy documentation across the industry remains deeply flawed. Stanford researchers found policies written in convoluted legal language and missing information users need to understand what happens to their data.
The Browser Extension Problem
The platforms themselves aren’t the only threat. In July 2025, Urban VPN Proxy - a Chrome extension with over six million users and a Google “Featured” badge - shipped an update that silently intercepted every conversation users had with eight AI assistants: ChatGPT, Claude, Gemini, Copilot, Perplexity, DeepSeek, Grok, and Meta AI.
The extension sent the captured conversations to Urban VPN’s servers, where they were sold to advertisers. The harvesting was enabled by default through hardcoded flags with no way to disable it short of uninstalling. Seven other extensions from the same publisher had the same code, putting more than eight million users’ AI conversations up for sale.
The irony: several of these extensions marketed “AI protection” as a feature. Google eventually pulled the Chrome versions in December 2025, but the Edge versions remained available longer.
The Bright Spot: On-Device AI
Apple Intelligence stands out as the strongest privacy-first approach in the market. Processing happens on your device whenever possible. When a task needs more computing power, Apple’s Private Cloud Compute sends only the relevant data to Apple silicon servers, processes it, returns the result, and deletes everything. No data is stored, no data is accessible to Apple, and independent researchers can verify the system.
The trade-off is that Apple Intelligence is currently less capable than cloud-based competitors. But for users who care about privacy above all else, it’s the only major offering where on-device processing is the default rather than the exception.
How to Opt Out: Platform by Platform
Here are the controls reported as of February 2026:
ChatGPT: Settings > Data Controls > turn off “Improve the model for everyone.” Takes effect immediately. Your conversations will still be stored for abuse monitoring but won’t feed into training.
Claude: Open Privacy Settings and turn off “You can help improve Claude.” Anthropic said Free, Pro, and Max users had until September 28, 2025 to accept the updated terms; the choice applied immediately to new and resumed chats or coding sessions. Work, Gov, Education, and API users were not affected.
Gemini: Go to your Gemini Apps Activity page and toggle it off. Also available: a “Temporary Chat” mode where conversations aren’t saved at all.
Copilot: For Microsoft 365 enterprise users, admins control data policies. For consumer Copilot, check Settings > Privacy and review what data sharing is enabled.
Grok: Settings > Privacy > opt out of training.
Perplexity: No clear mechanism for personal users to fully opt out of training. Enterprise customers are exempt.
DeepSeek: No opt-out mechanism. The only way to protect your data is to not use the consumer app and instead self-host the open-weight model locally.
Meta AI: No opt-out for AI training data. EU users can submit objection forms under GDPR, but the process is burdensome and outcomes are uncertain.
What You Can Do Right Now
1. Audit your settings today. Open each AI app you use and check whether training data sharing is enabled. If you haven’t touched these settings, it’s almost certainly on.
2. Use temporary or incognito modes. Claude, Gemini, and Perplexity all offer conversation modes where nothing is saved. Use them for anything sensitive.
3. Check your browser extensions. Any extension with access to your browser can read your AI conversations. Audit what you have installed, remove anything you don’t actively need, and be especially wary of VPN and ad-blocking extensions that request broad page access.
4. Consider local alternatives. Running models locally through tools like Ollama means your conversations never leave your machine. The latest open-weight models are competitive for many tasks. We published a guide to self-hosting your own ChatGPT alternative earlier today.
5. Separate personal from professional. Don’t paste proprietary code, financial information, medical details, or personal identifiers into any cloud-based AI tool unless you’ve verified the data handling policy for your specific plan tier. Consumer and enterprise terms are often very different.
6. Watch for policy changes. AI companies update their privacy policies frequently, and the trend has been toward collecting more data, not less. Anthropic’s shift from opt-in to opt-out in late 2025 happened with minimal fanfare. Set a calendar reminder to re-check your settings quarterly.
Why This Matters
The AI industry is splitting into two camps. One sells privacy as the product - enterprise platforms with contractual guarantees, audit trails, and compliance certifications. The other sells you as the product - free tiers funded by your data, conversations mined for training, and increasingly, ads based on what you discuss with a chatbot.
The gap between these two approaches is widening. For many people using free versions of these tools, default settings can route conversations, questions, and creative work into model-training systems unless users change the available controls. Colorado’s new Algorithmic Accountability Law, effective this month, is one of the first US laws to give consumers rights to notice, explanation, and appeal when AI systems make decisions about them. But federal privacy legislation still hasn’t caught up.
Until it does, the burden falls on you. Check your settings.